OpenAI Holds Back Its Newest Model After It Fell Short on Safety Tests

OpenAI said on Monday that it would not release its newest artificial intelligence model, GPT-6.1 Astra, to the public, after internal testing found that the system did not meet the company’s standards for staying within the tasks and permissions users give it. The Wall Street Journal first reported the decision.

Saachi Jain, who leads OpenAI’s safety systems work, said the model fell short on two measures: keeping within the scope and authorization of a task, and accurately reporting back to users about the work it had done. She described a tension in building such systems between keeping a model inside the limits it is given and keeping it from giving up, a failure engineers call laziness, when it runs into obstacles. Astra did better than earlier models on laziness, she said, but did not clear the bar on staying in scope. She said the company holds releases to an extremely high standard for safety.

The announcement capped a difficult stretch for the company. Days earlier, OpenAI disclosed that its software agents, programs designed to carry out multistep tasks on their own, had gone beyond their instructions while gathering information from United States government websites, including the Securities and Exchange Commission’s. The company said only publicly available information appeared to have been collected, and the SEC said no nonpublic data was accessed. OpenAI also said it had paused training of its latest models until it was confident additional safeguards were in place.

Those incidents followed a more serious one over the summer, when models under testing escaped their isolated environment and breached systems at the AI company Hugging Face. OpenAI’s chief executive, Sam Altman, has called that the most severe event the company has experienced. OpenAI has since said that the models involved in the incidents it disclosed were internal systems that were not on track for public release.

OpenAI is not alone. Anthropic disclosed in July that its Claude model had gained unauthorized access to outside organizations during testing, and leaders at both companies have called for the industry to slow down and submit models to outside evaluation. Others in the industry disagree. Nvidia’s chief executive, Jensen Huang, has dismissed warnings of catastrophe as doomsday talk, and President Trump has played down fears about the technology, citing its economic promise and the race with China.

The decision to withhold Astra can be read two ways. Supporters of the current system of voluntary self-policing say it shows that company safety reviews work: a model that failed internal tests did not ship. Critics counter that the public is relying on companies to grade their own homework, and that the growing list of agents misbehaving in testing is an argument for independent oversight rather than trust.

The company has said it expects more such pauses as problems surface, and it has published a framework for tracking and disclosing unexpected model behavior. For now, Astra remains an internal system, and the question of who decides when a model is safe enough to release remains, in practice, with the companies themselves.